It seems that a lot of people here are pesimistic.
Maybe you forgot that unified shader means shaders can be utilized as needed, wich means NO MORE stalling.
And believe me, effects can be done much much faster & efficient, 6-8 times faster isn't just random marketing **** for noobs. (MS's gotta pay me for saying this)
Oh you want an example?
Cube Mapping in DX10. In DX9 generating 1 cubemap requires 6 passes (for all 6 sides), while in DX10 all 6 sides of the cubemap can be generated in only 1 pass, with the use GS and render target array. (MS CubemapGS sample)
Imagine the speed gain if you have many objects to be cubemapped.
And the DX10 skinning I told you about in previous post, stream-out can really up the fps on multiple pass rendering. (MS skinning10 sample)
One traditional problem with GPU skinning was that if multiple passes needed to be performed on an object, the object needed to be skinned once for each pass. Take, for example, a scene with 3 shadow mapped light sources and a skinned character. To generate shadows for each of the lights the character needs to be drawn 3 times, once from the point of view of each light. (NOTE: Amplifying and streaming geometry to multiple render targets using the GS could also help aleviate this problem.) This would require that the Vertex Shader skin the character 3 times. The unfortunate reality is that this is two times too many. We really want to skin the character once and save this transformed character to use for the remainder of the frame. That way, if we had a scene that required 10 passes, the character wouldn't need to be skinned 10 times.
Fortunately, Stream Out allows us to do just that. Direct3D 10 allows the sample to skin the character once and stream the results out to a buffer in video memory. Then, for the remainder of the passes, the application just uses the pre-skinned vertices. For a character that may take 10 passes to render, this is 9-fold savings in vertex processing.
This sample does not render multiple passes, but instead renders the character multiple times. It's the same problem, just stated differently. When Stream Out is disabled, the application skins the character each time it renders it. However, when Stream Out is enabled, the skins the character once, saves the results via Stream Out, and renders the multiple characters using the pre-skinned geometry.
^ actual, real proof; denial would be just stupid.
Not just cubemapping or skinning, everything will be faster if the programmer can well utilize the new features of DX10.
I believe Crytek is really damn good at this.